An intelligent real-time nuclear radiation dose monitoring system based on deep learning

Abstract

The present invention discloses an intelligent real-time nuclear radiation dose monitoring system based on deep learning, comprising S1, a multimodal perception module for outputting radiation intensity data, environmental image data, and geographic location information with timestamps; S2, a preprocessing module for outputting structured feature sequences; S3, a fusion modeling module for generating a unified fusion feature vector; S4, an edge-end inference module for processing the fusion feature vector in stages and generating nuclear radiation dose prediction results; S5, a task scheduling module for dynamically determining whether inference tasks are executed on the edge or on the end node; S6, an anomaly detection module for determining whether there are abnormal changes in nuclear radiation dose; and S7, a system control module for coordinating the above modules. The present invention has the advantages of strong environmental adaptability, fast inference response, and high anomaly recognition accuracy.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to an intelligent real-time nuclear radiation dose monitoring system based on deep learning. Background Art

[0002] With the continued expansion of applications in nuclear energy technology, radiation medicine, nuclear facility decommissioning, and nuclear accident emergency response, dynamic perception and intelligent management of radiation doses in complex environments are placing higher demands on real-time performance and precision. In nuclear radiation scenarios, how to rapidly identify sudden dose anomalies, comprehensively model multimodal environmental changes, and efficiently schedule the inference process have become key issues that radiation monitoring systems urgently need to address.

[0003] In existing technologies, nuclear radiation dose monitoring mostly relies on data collection from a single sensor and an alarm strategy with a fixed threshold. Typical systems use Geiger counters, ionization chambers, or scintillators as radiation detectors, and perform data processing with simple filtering and judgment logic. This makes it difficult to reflect the dose evolution process under the influence of multiple sources in a complex radiation environment. At the same time, traditional systems lack the ability to integrate environmental images and spatial location information, and are unable to accurately identify regional risk boundaries in dynamic scenarios such as heterogeneous radiation distribution, mobile source interference, or sudden leaks. In addition, current methods generally adopt a centralized reasoning architecture, and the system response is highly dependent on network channels and back-end processing capabilities. It is difficult to adapt to the edge collaboration requirements in scenarios with multiple devices, weak connections, or delay-sensitive scenarios. The reasoning results have hidden dangers such as response lag and control link failure.

[0004] In the area of ​​dose anomaly detection and early warning, mainstream methods are still based on static threshold judgments and lack deep learning-driven adaptive recognition mechanisms. This makes it difficult to capture potential pattern changes in dose data and inadequately identify nonlinear growth, slow drift, or latent anomalies. This often results in the system failing to issue effective early warnings in the early stages of high-risk evolution, seriously impacting nuclear safety and control capabilities. Furthermore, some research has attempted to introduce neural networks for dose regression or radiation source identification, but these approaches generally remain at the model-building stage, lacking system integration with multimodal input, inference path switching, and anomaly handling processes. This makes it difficult to achieve a closed-loop intelligent response system from perception, modeling, inference, to alerting.

[0005] In summary, there is an urgent need for an intelligent real-time nuclear radiation dose monitoring system that integrates multimodal perception, edge-to-edge collaborative reasoning, and deep anomaly recognition mechanism to solve the core problems existing in current nuclear radiation monitoring, such as low environmental perception accuracy, untimely reasoning response, and insufficient ability to judge dose anomalies. Summary of the Invention

[0006] One purpose of the present invention is to propose an intelligent real-time monitoring system for nuclear radiation dose based on deep learning, integrate a multimodal perception modeling method of radiation intensity, environmental images and geographic location information, construct an inference architecture that collaborates with a lightweight model on the end and a deep edge model, and introduce an anomaly detection mechanism based on reconstruction error to realize dose change trend identification and multi-level early warning control. The inference path is dynamically adjusted through the task scheduling mechanism and intelligent response is completed in combination with time series data features. It has the advantages of strong environmental adaptability, fast inference response and high anomaly recognition accuracy.

[0007] According to an embodiment of the present invention, an intelligent real-time nuclear radiation dose monitoring system based on deep learning includes:

[0008] S1, multimodal perception module, used to collect nuclear radiation intensity data, environmental image data and geographic spatial location information, and output radiation intensity data, environmental image data and geographic spatial location information with timestamps;

[0009] S2, preprocessing module, is used to normalize the nuclear radiation intensity data, perform multi-scale feature extraction on the environmental image data, perform trajectory normalization on the geographic space location information, and output a structured feature sequence;

[0010] S3, fusion modeling module, is used to input the structured feature sequence into the deep fusion network based on the attention mechanism to generate a unified fusion feature vector;

[0011] S4, the edge inference module, includes a lightweight neural network model deployed on the edge and a deep neural network model deployed on the edge node. It is used to process the fused feature vectors in stages and generate nuclear radiation dose prediction results.

[0012] S5, the task scheduling module, is used to dynamically determine whether the inference task is executed on the lightweight neural network model on the client side or the deep neural network model on the edge node based on the current network status and device load;

[0013] S6, an anomaly detection module, used to execute an anomaly recognition model based on reconstruction error on the nuclear radiation dose prediction results to determine whether there is an abnormal change in nuclear radiation dose;

[0014] S7, system control module, is used to coordinate the communication, data transmission and status scheduling among the modules to achieve full process control.

[0015] Optionally, the S3 specifically includes:

[0016] S31, a modal feature alignment unit, configured to map the nuclear radiation intensity feature, the environmental image feature, and the geographic space location information feature in the structured feature sequence to a unified feature dimension, and generate an initial modal feature set {Fr ,F i ,F g}, where F r represents the uniformly coded nuclear radiation intensity feature vector, F i represents the feature vector of the environment image after processing by the convolutional network, F g Represents the feature vector of geographic spatial location information after spatial encoding;

[0017] S32, an attention weighted encoding unit, configured to introduce a cross-modal attention mechanism between the initial modal feature sets and calculate the attention allocation weights between the modalities:

[0018]

[0019] Among them, α m,n Represents the modal feature F m For modal characteristics F n The attention allocation weight, m, n∈{r,i,g}, f(F m ,F n ) is a correlation scoring function based on dot product, which is used to measure the feature dependency between modalities;

[0020] S33, a fusion representation generation unit, configured to perform a weighted combination of all modal features based on the attention allocation weights to generate a unified fusion feature vector F fusion , which is calculated as follows:

[0021]

[0022] S34, a regularization processing unit, for processing the fused feature vector F fusion Apply layer normalization and non-linear activation function operations.

[0023] Optionally, the S4 specifically includes:

[0024] S41, end-side inference unit, used to deploy lightweight neural network model M e , receiving the fusion feature vector F output by the fusion modeling module fusion , and generate the intermediate feature representation H based on the fused feature vector e , the M e It includes multi-layer linear mapping and normalization modules, and the calculation relationship is:

[0025] H e =M e (F fusion )=σ(W e ·F fusion +b e )

[0026] Among them, W e is the weight matrix, b e is the bias term, σ(·) is the nonlinear activation function;

[0027] S42, an intermediate feature buffer unit, configured to store the intermediate feature representation H e Stored in the local cache area, and determined whether to transmit uplink according to the inference allocation instruction issued by the system control module;

[0028] S43, edge node inference unit, used to deploy deep neural network model M c , after receiving the intermediate feature representation H e Then execute the subsequent reasoning process to generate the final nuclear radiation dose prediction value D c :

[0029] D c =M c (H e )

[0030] Among them, M c Contains multiple convolution, residual connection and attention mechanism modules to extract deep correlation features;

[0031] S44, the reasoning stage switching control unit is used to control the reasoning process in the lightweight neural network model M according to the real-time allocation result returned by the system task scheduling module. e With the deep neural network model M c Switch between and mark the current reasoning state;

[0032] S45, a model parameter synchronization unit, configured to periodically check the model version number and parameter verification code on the end-side and edge nodes, and complete the weight data consistency update through the system control module;

[0033] S46, an abnormal interruption processing unit, is used to automatically trigger the edge node to take over the complete inference task process when the end-side inference process or data transmission is interrupted, and return the prediction status label to the system control module after recovery.

[0034] Optionally, the S41 specifically includes:

[0035] S411, input feature receiving unit, used to receive the fusion feature vector output by the fusion modeling module Among them, d represents the feature dimension, represents the field of real numbers;

[0036] S412, a linear mapping processing unit, configured to input the fused feature vector into a lightweight neural network model M e , the lightweight neural network model consists of multiple linear mapping layers, where the structure of the lth layer mapping is:

[0037] Z (l) =W (l) ·H (l-1) +b (l)

[0038] Among them, W (l) represents the weight matrix of the lth layer, b (l) represents the bias vector, H (l-1) is the output of the previous layer, and the initial input satisfies H (0) =F fusion ;

[0039] S413, an activation and normalization unit, configured to perform nonlinear activation and normalization processing after the partial linear mapping layer;

[0040] S414, an output generation unit, configured to output an intermediate feature representation H at the end of the model. e =H (L) , where L represents the total number of layers of the neural network model, and H e Send to the intermediate feature buffer unit.

[0041] Optionally, the S43 specifically includes:

[0042] S431, edge feature receiving unit, used to receive the intermediate feature representation from the end side Among them, d ' Represents the intermediate feature dimension and feeds it into the deep neural network model M of the edge node c middle;

[0043] S432, convolution feature extraction unit, used to represent the intermediate feature H e The initial spatial dimension is expanded, and local context features are extracted through multiple two-dimensional convolutional layers. The convolution calculation structure is:

[0044] F (l) =Conv(F (l-1) ),F (0) =H e

[0045] Among them, Conv(·) represents the standard convolution operation, F (l) is the convolution output of the lth layer;

[0046] S433, residual connection unit, used to introduce a residual skip connection structure between the two-dimensional convolutional layers to establish a direct mapping relationship between input and output, satisfying F (l) =F (l) +F (l-1) , to enhance the cross-layer information retention capability, where F (l-1) is the convolution output of the previous layer;

[0047] S434, an attention enhancement unit, configured to embed a channel attention mechanism in a number of designated layers and perform adaptive weighted reconstruction on the channel dimension, wherein the attention weight is generated by normalizing the input features through a fully connected mapping;

[0048] S435, output generation unit, used to generate the final nuclear radiation dose prediction value D at the end of the deep neural network model c and transmits the predicted nuclear radiation dose value to the anomaly detection module and the system control module.

[0049] Optionally, the S44 specifically includes:

[0050] S441, the reasoning state perception unit is used to receive the real-time allocation result returned by the task scheduling module, the real-time allocation result is a scheduling state data packet, and the scheduling state data packet contains the network bandwidth B t , idle computing power on the terminal side C e and the real-time requirement of the task τ;

[0051] S442, the reasoning switching decision unit is used to construct the reasoning selection function φ(B t ,C e ,τ), where the function satisfies:

[0052]

[0053] Among them, θ1, θ2 are the preset computing resources and bandwidth thresholds, and φ outputs the model path that should be executed currently;

[0054] S443, an inference path execution unit, configured to select a function φ(B t ,C e ,τ) results control the fusion feature vector F fusion Distribute to the corresponding lightweight neural network model and deep neural network model, and simultaneously generate the path label λ for this reasoning round t ∈{e,c}, where e represents end-side reasoning and c represents edge reasoning;

[0055] S444, a state tag management unit is used to set the path tag λ t It is recorded together with the timestamp, task number and prediction result identifier to form a structured reasoning status log, which is periodically synchronized to the system control module.

[0056] Optionally, the S6 specifically includes:

[0057] S61, prediction result cache unit, used to receive and temporarily store the nuclear radiation dose prediction value D generated by the edge node reasoning unit c, and store them in chronological order to form a sequence sample set Where T is the length of the sample time window;

[0058] S62, a reconstruction model inference unit is used to input the sequence sample set into a preset anomaly recognition model R(·) to generate a reconstructed sequence in, Represents the reconstructed sequence result of the t-th prediction value;

[0059] S63, an error calculation and determination unit, configured to calculate a reconstruction error sequence E={e 1 ,e 2 ,…,e T}, where the error value is defined as:

[0060]

[0061] And compare the error value with the set threshold δ, if there is e T >δ, then mark the moment as an abnormal point;

[0062] S64, an abnormality label generating unit, used to number the moment marked as abnormal {t a}Convert to abnormal event label set And send the collection to the system control module.

[0063] The beneficial effects of the present invention are:

[0064] (1) The present invention achieves joint modeling of dose distribution in complex radiation environments by constructing a multimodal perception mechanism that integrates radiation intensity, environmental images, and geographic spatial information. This breaks through the limitations of traditional reliance on single sensor data and improves the system's environmental perception stability and nuclear radiation dose measurement accuracy under conditions of spatial heterogeneity, target occlusion, and sudden interference.

[0065] (2) The present invention designs an edge-to-end collaborative reasoning architecture, introduces a hierarchical reasoning process of lightweight neural network models and deep neural network models, and dynamically selects the reasoning path through the resource scheduling function, effectively reducing the system's dependence on central node computing and communication, adapting to real-time response requirements under different network conditions, and enhancing flexibility and processing efficiency in edge deployment scenarios.

[0066] (3) The present invention constructs a deep anomaly recognition mechanism based on reconstruction error, reconstructs and analyzes the difference of nuclear radiation dose prediction results, and generates high-frequency dynamic anomaly labels in combination with the threshold judgment strategy. Compared with the fixed threshold alarm mechanism, it can identify hidden, continuous and slowly rising dose anomaly changes, thereby enhancing the system's detection sensitivity and early warning capabilities in the initial stage of nuclear accidents.

[0067] (4) The present invention proposes a linkage mechanism between reasoning state marking and path labeling, establishes a unified reasoning state log management model, and realizes structured tracking management of the entire process from feature input, model call to prediction output. It has high auditability and model operation closed-loop guarantee capabilities, providing basic data support for subsequent security audits and operation and maintenance diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0069] Figure 1 This is the overall structural framework diagram of the intelligent real-time nuclear radiation dose monitoring system based on deep learning proposed in this invention. DETAILED DESCRIPTION

[0070] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0071] refer to Figure 1 , an intelligent real-time nuclear radiation dose monitoring system based on deep learning, the system comprising:

[0072] S1, multimodal perception module, used to collect nuclear radiation intensity data, environmental image data and geographic spatial location information, and output radiation intensity data, environmental image data and geographic spatial location information with timestamps;

[0073] In the S1 multimodal perception module of the present invention, the collected nuclear radiation intensity data, environmental image data, and geospatial location information constitute the core input content of the system. Among them, nuclear radiation intensity data is collected by radiation detectors and includes dose rate values, pulse count values, sampling timestamps, and optional energy spectrum distribution information, which is used to reflect the radiation level and its changing trend in the current area. Environmental image data is acquired by visible light cameras or infrared imaging devices and includes RGB image frames, continuous image sequences, image metadata (such as shooting time, viewing angle, etc.), and infrared images used in specific tasks to capture structural features, lighting conditions, and abnormal events in the monitoring environment. Geospatial location information is provided by the GNSS positioning module and / or IMU component and includes three-dimensional coordinates (latitude, longitude, and altitude), continuous trajectory data, positioning accuracy indicators, and timestamps, etc., which are used to accurately calibrate the geographic location and movement path of the monitoring equipment. The above three types of data are all accompanied by a unified high-precision timestamp. Through a synchronization mechanism, they form a structured multi-source observation data stream, which serves as the standard input for subsequent preprocessing and model building processes, providing a comprehensive and synchronized perception foundation for spatial modeling, dynamic perception, and intelligent early warning of nuclear radiation dose.

[0074] In this embodiment, the S1 multimodal perception module includes a nuclear radiation intensity perception unit, an environmental image acquisition unit and a geospatial positioning unit. The nuclear radiation intensity perception unit is used to collect nuclear radiation dose-related data in the monitoring area, and obtains time series information representing changes in radiation levels by continuously sampling the surrounding radiation intensity. The environmental image acquisition unit is used to obtain image data under the current monitoring environment, and uses the onboard imaging device to take timed photos of the monitoring area to characterize the structure, occlusion, lighting or other potential influencing factors in the environment. The geospatial positioning unit is used to determine the spatial position of the perception device, and obtains the geographic coordinate information of the device through the positioning component to achieve a one-to-one correspondence between the perception data and the spatial position. After collection, the above-mentioned various types of perception data are time-aligned through the synchronization control module, and all data are accompanied by a unified timestamp to form a structured multi-source fusion data set. The multimodal perception module supports periodic acquisition and output, and serves as the input entrance of the entire system, providing a synchronous and complete perception information foundation for subsequent data preprocessing, feature extraction and model reasoning.

[0075] S2, preprocessing module, is used to normalize the nuclear radiation intensity data, perform multi-scale feature extraction on the environmental image data, perform trajectory normalization on the geographic space location information, and output a structured feature sequence;

[0076] In this embodiment, the S2 preprocessing module includes a radiation data normalization processing unit, an environmental image feature extraction unit, and a spatial trajectory normalization processing unit, which are used to perform format conversion and feature regularization on the data collected by the multimodal perception module to generate a structured multi-source input feature sequence. Among them, the radiation data normalization processing unit is used to scale the dose data collected by the nuclear radiation intensity perception unit, unify the data distribution output by different sensors, and reduce input inconsistency caused by equipment drift, environmental interference, or data fluctuations. The environmental image feature extraction unit is used to extract multi-level visual structure information from the collected image frames, perform multi-scale processing on the image data through the feature encoding module, and convert the original image into an image feature vector representing texture, edge, and shape, so as to facilitate the subsequent model to distinguish the environmental state. The spatial trajectory normalization processing unit is used to perform path regularization processing on the position data collected by the geospatial positioning unit. Through trajectory interpolation, coordinate remapping, and smoothing filtering, the spatial resolution and sampling density are unified to ensure that the spatial information is consistent with other modal data in terms of time and sequence structure. After processing, all types of data are converted into structured feature items in a unified format and organized into standardized feature sequences in chronological order, which serve as the input data set of the fusion modeling module, providing a standardized and neat input feature foundation for subsequent multimodal joint modeling.

[0077] S3, fusion modeling module, is used to input the structured feature sequence into the deep fusion network based on the attention mechanism to generate a unified fusion feature vector;

[0078] In this embodiment, S3 specifically includes:

[0079] S31, a modal feature alignment unit, configured to map the nuclear radiation intensity feature, the environmental image feature, and the geographic space location information feature in the structured feature sequence to a unified feature dimension, and generate an initial modal feature set {F r ,F i ,F g}, where F r represents the uniformly coded nuclear radiation intensity feature vector, F i represents the feature vector of the environment image after processing by the convolutional network, F g Represents the feature vector of geographic spatial location information after spatial encoding;

[0080] S32, an attention weighted encoding unit, configured to introduce a cross-modal attention mechanism between the initial modal feature sets and calculate the attention allocation weights between the modalities:

[0081]

[0082] Among them, α m,n Represents the modal feature Fm For modal characteristics F n The attention allocation weight, m, n∈{r,i,g}, f(F m ,F n ) is a correlation scoring function based on dot product, which is used to measure the feature dependency between modalities;

[0083] Specifically, for a given modal feature F m With F n , the system evaluates the correlation function f(F m ,F n ) Calculate the feature matching degree of the two and obtain the attention allocation weight α through the Softmax normalization operation m,n , indicating that in the current fusion process, the modality F m For mode F n The calculation method of this weight is shown in the formula, which ensures that the sum of all assigned weights in the same mode is 1, meeting the probability normalization requirement.

[0084] S33, a fusion representation generation unit, configured to perform a weighted combination of all modal features based on the attention allocation weights to generate a unified fusion feature vector F fusion , which is calculated as follows:

[0085]

[0086] The fusion process introduces an attention weighting mechanism on the modality dimension to perform an attention weighting on each modality feature F. n Assign a weight α determined by its relationship with other modalities m,n In the specific implementation, the system performs m The other modal features of interest are linearly weighted and summed across all modal dimensions to obtain the final fused feature representation. The above process, as shown in the formula, uses a dual summation structure to ensure that all interactions between modalities are included in the fusion calculation.

[0087] S34, a regularization processing unit, for processing the fused feature vector F fusion Apply layer normalization and non-linear activation function operations.

[0088] This embodiment realizes unified feature encoding and dynamic weighted fusion of nuclear radiation intensity, environmental images and geographic spatial information by constructing a modal feature alignment unit, an attention weighted encoding unit, a fusion representation generation unit and a regularization processing unit. It can adaptively adjust the information interaction weight according to the correlation between each modality, thereby effectively integrating the key features in multi-source heterogeneous data, improving the representation ability and discriminability of the fusion representation, providing a more robust and environmentally adaptable input basis for subsequent reasoning modules, and enhancing the system's intelligent recognition ability and processing accuracy in complex nuclear radiation scenarios.

[0089] S4, the edge inference module, includes a lightweight neural network model deployed on the edge and a deep neural network model deployed on the edge node. It is used to process the fused feature vectors in stages and generate nuclear radiation dose prediction results.

[0090] In this embodiment, S4 specifically includes:

[0091] S41, end-side inference unit, used to deploy lightweight neural network model M e , receiving the fusion feature vector F output by the fusion modeling module fusion , and generate the intermediate feature representation H based on the fused feature vector e , the M e It includes multi-layer linear mapping and normalization modules, and the calculation relationship is:

[0092] H e =M e (F fusion )=σ(W e ·F fusion +b e )

[0093] Among them, W e is the weight matrix, b e is the bias term, σ(·) is the nonlinear activation function;

[0094] This formula performs feature mapping on the input vector by combining linear transformation and nonlinear activation function. fusion First, through the linear mapping of the weight matrix and the translation adjustment of the bias term, a new feature space expression is formed; then, the activation function σ(·) is introduced to perform nonlinear processing on the mapping result to improve the model's ability to express complex patterns. The final H e Indicates the output result of the client-side inference module, which is passed as input to the edge node inference module or cache module.

[0095] S42, an intermediate feature buffer unit, configured to store the intermediate feature representation H eStored in the local cache area, and determined whether to transmit uplink according to the inference allocation instruction issued by the system control module;

[0096] S43, edge node inference unit, used to deploy deep neural network model M c , after receiving the intermediate feature representation H e Then execute the subsequent reasoning process to generate the final nuclear radiation dose prediction value D c :

[0097] D c =M c (H e )

[0098] Among them, M c Contains multiple convolution, residual connection and attention mechanism modules to extract deep correlation features;

[0099] The formula represents the depth model M c The forward calculation process, that is, H e As input, it passes through a multi-layer neural structure composed of convolution, residual connection and attention mechanism modules to extract information with more semantic depth and local context, and finally outputs the dose estimation result D in scalar or vector form. c This structure can effectively integrate the capabilities of local feature enhancement and global dependency modeling, improving the system's accuracy in distinguishing complex radiation patterns.

[0100] In specific use, model M c It supports continuous operation on edge nodes with computing resources. When the system task scheduling module decides to transfer the inference task to edge processing, the deep prediction process of nuclear radiation dose is completed in the manner defined by this formula. c It will be synchronously transmitted to the anomaly detection module and system control module for subsequent risk identification and response decisions.

[0101] S44, the reasoning stage switching control unit is used to control the reasoning process in the lightweight neural network model M according to the real-time allocation result returned by the system task scheduling module. e With the deep neural network model M c Switch between and mark the current reasoning state;

[0102] S45, a model parameter synchronization unit, configured to periodically check the model version number and parameter verification code on the end-side and edge nodes, and complete the weight data consistency update through the system control module;

[0103] S46, an abnormal interruption processing unit, is used to automatically trigger the edge node to take over the complete inference task process when the end-side inference process or data transmission is interrupted, and return the prediction status label to the system control module after recovery.

[0104] The S41 specifically includes:

[0105] S411, input feature receiving unit, used to receive the fusion feature vector output by the fusion modeling module Among them, d represents the feature dimension, represents the field of real numbers;

[0106] S412, a linear mapping processing unit, configured to input the fused feature vector into a lightweight neural network model M e , the lightweight neural network model consists of multiple linear mapping layers, where the structure of the lth layer mapping is:

[0107] Z (l) =W (l) ·H (l-1) +b (l)

[0108] Among them, W (l) represents the weight matrix of the lth layer, b (l) represents the bias vector, H (l-1) is the output of the previous layer, and the initial input satisfies H (0) =F fusion ;

[0109] The formula represents the linear mapping structure of the first layer in the lightweight model, specifically: the output vector H of the previous layer (l-1) With the current layer weight matrix W (l) Multiply and add the bias term b (l) , forming the output Z of the current layer (l) This process is repeated in each layer of the model to extract the linear change pattern and dimensionality transformation rules in the input features.

[0110] In the model initialization stage, the input vector satisfies H (0) =F fusion , that is, directly receiving the fusion feature representation generated by the fusion modeling module, and then each layer of linear mapping unit performs feature mapping in turn based on this input.

[0111] S413, an activation and normalization unit, configured to perform nonlinear activation and normalization processing after the partial linear mapping layer;

[0112] S414, an output generation unit, configured to output an intermediate feature representation H at the end of the model. e =H (L) , where L represents the total number of layers of the neural network model, and H e Send to the intermediate feature buffer unit.

[0113] The S43 specifically includes:

[0114] S431, edge feature receiving unit, used to receive the intermediate feature representation from the end side Among them, d ' Represents the intermediate feature dimension and feeds it into the deep neural network model M of the edge node c middle;

[0115] S432, convolution feature extraction unit, used to represent the intermediate feature H e The initial spatial dimension is expanded, and local context features are extracted through multiple two-dimensional convolutional layers. The convolution calculation structure is:

[0116] F (l) =Conv(F (l-1) )

[0117] F (0) =H e

[0118] Among them, Conv(·) represents the standard convolution operation, F (l) is the convolution output of the lth layer;

[0119] The formula defines the standard computational structure of the first layer in a convolutional network, which is to take the convolution output F of the previous layer as (l-1) As the input of the current convolutional layer, it performs feature extraction operation on it through the convolution kernel and outputs the feature representation F of the current layer. (l) The whole process starts from the initial input F (0) =H e Start by iterating and expanding layer by layer.

[0120] The Conv(·) operation in the formula represents a standard two-dimensional convolution operation, which scans the input feature map in the form of a sliding window to extract regional pattern information. This operation can set different convolution kernel sizes, strides, and number of channels at each layer to meet the needs of multi-scale feature extraction.

[0121] S433, residual connection unit, used to introduce a residual skip connection structure between the two-dimensional convolutional layers to establish a direct mapping relationship between input and output, satisfying F (l) =F (l) +F (l-1) , to enhance the cross-layer information retention capability, where F (l-1) is the convolution output of the previous layer;

[0122] S434, an attention enhancement unit, configured to embed a channel attention mechanism in a number of designated layers and perform adaptive weighted reconstruction on the channel dimension, wherein the attention weight is generated by normalizing the input features through a fully connected mapping;

[0123] S435, output generation unit, used to generate the final nuclear radiation dose prediction value D at the end of the deep neural network model c and transmits the predicted nuclear radiation dose value to the anomaly detection module and the system control module.

[0124] The S44 specifically includes:

[0125] S441, the reasoning state perception unit is used to receive the real-time allocation result returned by the task scheduling module, the real-time allocation result is a scheduling state data packet, and the scheduling state data packet contains the network bandwidth B t , idle computing power on the terminal side C e and the real-time requirement of the task τ;

[0126] S442, the reasoning switching decision unit is used to construct the reasoning selection function φ(B t ,C e ,τ), where the function satisfies:

[0127]

[0128] Among them, θ1, θ2 are the preset computing resources and bandwidth thresholds, and φ outputs the model path that should be executed currently;

[0129] The function φ defined in the formula represents a conditional control strategy. When the remaining computing resources C on the end side e is greater than or equal to the preset threshold θ1, and the current network bandwidth B t When the threshold value θ2 is lower than the limit, the lightweight neural network model M is preferred. e Execute the reasoning process; otherwise, the system automatically switches to the deep neural network model M of the edge node c To complete the remaining calculation tasks.

[0130] S443, an inference path execution unit, configured to select a function φ(B t ,C e ,τ) results control the fusion feature vector F fusion Distribute to the corresponding lightweight neural network model and deep neural network model, and simultaneously generate the path label λ for this reasoning round t ∈{e,c}, where e represents end-side reasoning and c represents edge reasoning;

[0131] S444, a state tag management unit is used to set the path tag λ t It is recorded together with the timestamp, task number and prediction result identifier to form a structured reasoning status log, which is periodically synchronized to the system control module.

[0132] This implementation implements a phased processing mechanism for fused feature vectors by constructing a collaborative reasoning architecture between a lightweight neural network model on the end and a deep neural network model on the edge. This not only quickly generates intermediate feature representations on the end to meet the initial response requirements in delay-sensitive scenarios, but also further completes deep feature mining and dose prediction at the edge nodes, supporting the reasoning capabilities of complex models. At the same time, the system ensures the consistency and stability of reasoning between different computing levels through inference path switching control and model parameter synchronization mechanisms, effectively balancing computing efficiency and prediction accuracy, and enhancing the adaptability and continuous operation capabilities of the nuclear radiation monitoring system under multi-scenario deployment.

[0133] S5, the task scheduling module, is used to dynamically determine whether the inference task is executed on the lightweight neural network model on the client side or the deep neural network model on the edge node based on the current network status and device load;

[0134] In this embodiment, the S5 task scheduling module includes a state information acquisition unit, a scheduling decision generation unit and a scheduling instruction broadcasting unit, which are used to dynamically control the execution position of the inference task and realize the intelligent allocation of the model path.

[0135] The status information collection unit is used to continuously receive operating status data from the end-side and edge nodes. The collected content includes but is not limited to scheduling-related parameters such as the current available computing power, memory load, system idle rate, and current network channel bandwidth of the end-side device.

[0136] The scheduling decision generation unit constructs a task scheduling logic function based on the currently collected status data. By comparing it with a preset threshold, it determines whether the inference task is capable of being completed on the edge. If the edge resources meet the requirements for lightweight inference and the network status is not suitable for data uplink, the system prioritizes completing the inference process locally. Otherwise, the system offloads the instruction inference task to the edge node.

[0137] The scheduling instruction broadcast unit is used to send the decision results in the form of control instructions to the inference path execution module and update the task path record table of the system control module to ensure that subsequent modules receive and process data under the correct path.

[0138] This implementation introduces a task scheduling module based on resource perception and bandwidth judgment to implement a flexible switching control mechanism for inference tasks between the terminal side and the edge node. It can adjust the model execution location in real time according to the system operation status, taking into account both inference efficiency and resource utilization, avoiding performance degradation caused by single-node computing bottlenecks or network fluctuations, and enhancing the system's scalability, stability, and response speed in various scenarios.

[0139] S6, an anomaly detection module, used to execute an anomaly recognition model based on reconstruction error on the nuclear radiation dose prediction results to determine whether there is an abnormal change in nuclear radiation dose;

[0140] S61, prediction result cache unit, used to receive and temporarily store the nuclear radiation dose prediction value D generated by the edge node reasoning unit c , and store them in chronological order to form a sequence sample set Where T is the length of the sample time window;

[0141] S62, a reconstruction model inference unit is used to input the sequence sample set into a preset anomaly recognition model R(·) to generate a reconstructed sequence in, Represents the reconstructed sequence result of the t-th prediction value;

[0142] S63, an error calculation and determination unit, configured to calculate a reconstruction error sequence E={e 1 ,e 2 ,…,e T}, where the error value is defined as:

[0143]

[0144] And compare the error value with the set threshold δ, if there is e T >δ, then mark the moment as an abnormal point;

[0145] Specifically, the error value calculation formula measures the difference between the predicted value and the reconstructed value by the Euclidean distance (L2 norm). This error can reflect whether the prediction result at the current moment deviates from the normal mode. t When the set threshold δ is exceeded, the system will mark the moment as an abnormal point.

[0146] S64, an abnormality label generating unit, used to number the moment marked as abnormal {t a}Convert to abnormal event label set And send the collection to the system control module.

[0147] This implementation method constructs an inference path selection function based on real-time scheduling status to achieve a comprehensive judgment of the current network bandwidth, end-side computing power and task real-time requirements. It can dynamically select end-side lightweight models or edge deep models for inference task allocation, which not only improves the resource adaptability of the model in multiple scenarios, but also enhances the stability and response flexibility of the system under network fluctuations or equipment load changes. It effectively avoids delay anomalies or resource waste caused by fixed inference paths, and ensures the high availability and processing efficiency of the nuclear radiation monitoring system in edge-end collaborative scenarios.

[0148] S7, system control module, is used to coordinate the communication, data transmission and status scheduling among the modules to achieve full process control.

[0149] In this embodiment, the S7 system control module includes a communication coordination unit, a data synchronization management unit and a state scheduling control unit, which are used to uniformly manage the interaction process between various functional modules within the system to ensure the stability and orderliness of the multi-stage processing process.

[0150] The communication coordination unit establishes data communication channels between modules, ensuring low-latency and low-packet-loss information transfer between the client-side inference module, edge inference module, task scheduling module, and anomaly detection module. This unit supports message queuing and transmission status monitoring, automatically adjusting packet size and transmission frequency based on network conditions.

[0151] The data synchronization management unit is used to organize and cache key data streams such as multimodal perception data, model output results and prediction labels, and to time-series mark processing tasks in different time periods to ensure the consistency and traceability of data in processes such as reasoning, anomaly identification and warning output.

[0152] The state scheduling control unit is used to coordinate model switching, module scheduling and task instruction broadcasting functions. According to the information feedback from the task scheduling module and the system status log, it dynamically adjusts the module execution priority and operation status, and periodically transmits the overall system operation status back to the upper platform interface.

[0153] Through the system control module, all sub-modules can complete the closed-loop operation process of "acquisition-modeling-inference-identification-early warning" under a unified scheduling system, realizing the coordinated control of the entire process of task scheduling, data flow and status maintenance.

[0154] This implementation method introduces a system control module to centrally manage communications, data, and scheduling tasks, achieving process docking and control status consistency between functional units, and improving the overall operational stability and module collaboration capabilities of the system. At the same time, based on the module status monitoring and log feedback mechanism, the inference path, model output, and abnormal events can be fully recorded and tracked, enhancing the system's manageability, maintainability, and adaptability to complex working conditions, providing strong system-level protection for nuclear radiation monitoring tasks.

[0155] Example:

[0156] To verify the feasibility of this invention, the proposed intelligent real-time nuclear radiation dose monitoring system based on deep learning was applied to a mobile radiation monitoring exercise conducted by a national nuclear emergency response technical support center in coastal Shandong. This exercise simulated a sudden radiation leak from a nuclear facility and required the monitoring system to perform high-frequency, low-latency, and intelligent sensing and analysis of nuclear radiation doses in a dynamic mobile environment. It also automatically issued warning signals in the event of an anomaly, enabling on-site decision support and remote collaborative collaboration.

[0157] The scenario for this exercise involved a nuclear facility failure with a suspected risk of radioactive material spread. Emergency response units were required to rapidly deploy an unmanned monitoring fleet within a 5-kilometer radius of the incident to collect and analyze data in real time. Due to factors such as dense construction, complex traffic conditions, and uneven signal coverage within the area, traditional static point-based monitoring systems struggled to meet the demands for rapid response and high-precision identification. Furthermore, existing equipment was unable to effectively process the real-time fusion modeling of multimodal data, including images, location information, and nuclear radiation intensity. This severely limited the efficiency of on-site emergency monitoring and the accuracy of judgments.

[0158] In this scenario, the system described in the present invention operates jointly through the end-side perception and reasoning modules deployed on 6 monitoring vehicles and the edge computing nodes set up along the road. The monitoring vehicle is equipped with a multimodal perception unit, which periodically collects information including nuclear radiation dose, surrounding environment images and GNSS positioning information, and uses the task scheduling module to determine whether the data is processed by the end-side model or uploaded to the edge node for inference. After receiving different modal features, the system's built-in deep fusion module uses the attention mechanism for weighted modeling, extracts stable feature vectors and inputs them into the neural network model. For possible abnormal dose evolution trends, the system uses a reconstruction error strategy for intelligent identification, automatically marks abnormal time periods and issues multi-level warning information, which is synchronized to the mobile command platform in real time.

[0159] During the actual test, the drill lasted four hours, covering a monitoring radius of approximately 8.7 kilometers, and the system collected over 32,000 sets of time series data. The system accurately detected three simulated radiation anomalies and responded within 12 seconds of each significant radiation level fluctuation. To clearly demonstrate the actual results before and after deployment, the following data table provides specific data.

[0160] Table 1: Statistics of nuclear radiation anomaly detection drill results based on the system of the present invention

[0161]

[0162] As can be seen from Table 1, during the actual drill, the system deployed by the present invention demonstrated significant advantages in the entire process of data collection, reasoning and judgment, and response feedback. In particular, in terms of the response speed of nuclear radiation anomaly detection, the average delay was reduced from 96 seconds in the traditional system to 11.7 seconds, greatly shortening the judgment and response time window in the early stages of the accident and improving the efficiency of emergency response. At the same time, under complex terrain and unstable network connection conditions, the system can dynamically adjust the reasoning execution path through the task scheduling module, reasonably allocate end-side reasoning resources, and significantly reduce the impact of end-side resource bottlenecks on judgment speed and accuracy. The overall fusion recognition accuracy of the system reached 97.6%, which is more than 15% higher than the traditional method.

[0163] In one specific scenario, the system successfully identified a simulated anomaly source hidden in a low-lying corner, difficult to detect using conventional means. The error between its predicted location and the actual source was only 7.5 meters, far below the error range of traditional early warning equipment. Furthermore, throughout the data collection process, the system automatically annotated over 90% of the time periods for anomalies and, in conjunction with a multi-channel early warning system, immediately transmitted event information to the remote control platform, ensuring the real-time and full coverage of the drill.

[0164] According to feedback from technicians and test reports, the deployment of this system has resulted in higher levels of automation, identification accuracy, and system stability for the entire nuclear radiation monitoring task. This system is particularly well-suited for scenarios requiring sudden environmental changes, unstable signal conditions, or heterogeneous multi-terminal sensing. The system not only supports flexible deployment and scalable maintenance, but also boasts complete closed-loop information processing and intelligent reasoning capabilities, significantly improving the efficiency of identifying unusual nuclear radiation events, the accuracy of judgments, and the timeliness of response decisions during nuclear emergency response.

[0165] In summary, the intelligent real-time nuclear radiation dose monitoring system based on deep learning proposed in the present invention has excellent performance in real environments, especially in data fusion, adaptive reasoning and abnormal intelligent detection, which has significant advantages over traditional systems. This fully demonstrates the practicality, advancement and scalability of this technical solution in complex radiation monitoring tasks.

[0166] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. An intelligent real-time monitoring system for nuclear radiation dose based on deep learning, characterized in that: The system comprises: S1, multimodal perception module, used to collect nuclear radiation intensity data, environmental image data and geographic spatial location information, and output radiation intensity data, environmental image data and geographic spatial location information with timestamps; S2, preprocessing module, is used to normalize the nuclear radiation intensity data, perform multi-scale feature extraction on the environmental image data, perform trajectory normalization on the geographic space location information, and output a structured feature sequence; S3, the fusion modeling module, is used to input the structured feature sequence into a deep fusion network based on the attention mechanism. The network introduces a cross-modal attention mechanism to dynamically calculate and assign the dependency weights between the features of each modality to generate a unified fusion feature vector; S4, the edge inference module, includes a lightweight neural network model deployed on the edge and a deep neural network model deployed on the edge node. It is used to process the fused feature vectors in stages and generate nuclear radiation dose prediction results. S5, the task scheduling module, is used to dynamically determine whether the inference task is executed on the lightweight neural network model on the client side or the deep neural network model on the edge node based on the current network status and device load; S6, an anomaly detection module, used to execute an anomaly recognition model based on reconstruction error on the nuclear radiation dose prediction results to determine whether there is an abnormal change in nuclear radiation dose; S7, system control module, used to coordinate the communication, data transmission and status scheduling between the modules to achieve full process control; The S4 specifically includes: S41, end-side inference unit, used to deploy lightweight neural network models , receiving the fusion feature vector output by the fusion modeling module , and generate intermediate feature representation based on the fused feature vector , It includes multi-layer linear mapping and normalization modules, and the calculation relationship is: ； in, is the weight matrix, is the bias term, is a nonlinear activation function; S42, intermediate feature buffer unit, used to represent the intermediate feature Stored in the local cache area, and determined whether to transmit uplink according to the inference allocation instruction issued by the system control module; S43, edge node inference unit, used to deploy deep neural network models , after receiving the intermediate feature representation Then execute the subsequent reasoning process to generate the final nuclear radiation dose prediction value : ； in, Contains multiple convolution, residual connection and attention mechanism modules to extract deep correlation features; S44, the inference phase switching control unit is used to control the inference process in the lightweight neural network model according to the real-time allocation results returned by the system task scheduling module and deep neural network models Switch between and mark the current reasoning state; S45, a model parameter synchronization unit, configured to periodically check the model version number and parameter verification code on the end-side and edge nodes, and complete the weight data consistency update through the system control module; S46, an abnormal interruption processing unit, is used to automatically trigger the edge node to take over the complete inference task process when the end-side inference process or data transmission is interrupted, and return the prediction status label to the system control module after recovery.

2. The intelligent real-time monitoring system for nuclear radiation dose based on deep learning according to claim 1 is characterized in that: The S3 specifically includes: S31, a modal feature alignment unit, for mapping the nuclear radiation intensity feature, the environmental image feature, and the geographic space location information feature in the structured feature sequence to a unified feature dimension, to generate an initial modal feature set ,in, represents the uniformly coded nuclear radiation intensity feature vector, represents the feature vector of the environment image after processing by the convolutional network, Represents the feature vector of geographic spatial location information after spatial encoding; S32, an attention weighted encoding unit, configured to introduce a cross-modal attention mechanism between the initial modal feature sets and calculate the attention allocation weights between the modalities: ； in, Representing modal characteristics Modal characteristics The attention allocation weight, , is a dot-product-based correlation scoring function used to measure the feature dependency between modalities. Represents a modal feature vector in the initial modal feature set; S33, a fusion representation generation unit, configured to perform a weighted combination of all modal features based on the attention allocation weights to generate a unified fusion feature vector , which is calculated as follows: ； S34, a regularization processing unit, for processing the fused feature vector Apply layer normalization and non-linear activation function operations.

3. The intelligent real-time monitoring system for nuclear radiation dose based on deep learning according to claim 1 is characterized in that: The S41 specifically includes: S411, input feature receiving unit, used to receive the fusion feature vector output by the fusion modeling module ,in, represents the feature dimension, represents the field of real numbers; S412, a linear mapping processing unit, configured to input the fused feature vector into a lightweight neural network model , the lightweight neural network model consists of multiple linear mapping layers, where the The structure of the layer map is: ； in, Indicates the The weight matrix of the layer, represents the bias vector, is the output of the previous layer, and the initial input satisfies ; S413, an activation and normalization unit, configured to perform nonlinear activation and normalization processing after the partial linear mapping layer; S414, output generation unit, used to output intermediate feature representation at the end of the model ,in, Represents the total number of layers of the neural network model, and Send to the intermediate feature buffer unit.

4. The intelligent real-time monitoring system for nuclear radiation dose based on deep learning according to claim 1 is characterized in that: The S43 specifically includes: S431, edge feature receiving unit, used to receive the intermediate feature representation from the end side ,in, Represents the intermediate feature dimension and feeds it into the deep neural network model of the edge node middle; S432, convolution feature extraction unit, used to represent the intermediate features The initial spatial dimension is expanded, and local context features are extracted through multiple two-dimensional convolutional layers. The convolution calculation structure is: ； in, represents the standard convolution operation, For the Layer convolution output; S433, residual connection unit, used to introduce a residual skip connection structure between the two-dimensional convolutional layers to establish a direct mapping relationship between input and output, satisfying , to enhance the cross-layer information retention capability, where is the convolution output of the previous layer; S434, attention enhancement unit, is used to embed the channel attention mechanism in several specified layers and perform adaptive weighted reconstruction on the channel dimension. The attention weight is generated by normalizing the input features through the fully connected mapping; S435, output generation unit, used to generate the final nuclear radiation dose prediction value at the end of the deep neural network model and transmits the predicted nuclear radiation dose value to the anomaly detection module and the system control module.

5. The intelligent real-time monitoring system for nuclear radiation dose based on deep learning according to claim 1 is characterized in that: The S44 specifically includes: S441, the reasoning state perception unit is used to receive the real-time allocation result returned by the task scheduling module, the real-time allocation result is a scheduling state data packet, and the scheduling state data packet contains the network bandwidth , idle computing power on the end side Real-time requirements of tasks ; S442, inference switching decision unit, used to build an inference selection function based on scheduling state data , where the function satisfies: ； in, To preset computing resources and bandwidth thresholds, To output the model path that should be executed currently; S443, an inference path execution unit, configured to select a function according to the inference The result controls the fusion feature vector Distribute to the corresponding lightweight neural network model and deep neural network model, and simultaneously generate the path label for this reasoning round ,in, Indicates client-side reasoning. represents marginal reasoning; S444, a status mark management unit is used to set the path label It is recorded together with the timestamp, task number and prediction result identifier to form a structured reasoning status log, which is periodically synchronized to the system control module.

6. The intelligent real-time monitoring system for nuclear radiation dose based on deep learning according to claim 3 is characterized in that: The S6 specifically includes: S61, prediction result cache unit, used to receive and temporarily store the nuclear radiation dose prediction value generated by the edge node reasoning unit , and store them in chronological order to form a sequence sample set ,in, is the length of the sample time window; S62, reconstruction model inference unit, used to input the sequence sample set into the preset anomaly recognition model , generating the reconstruction sequence ,in, Indicates the The reconstructed sequence results of the predicted values; S63, error calculation and determination unit, used to calculate the reconstruction error sequence based on the difference between the nuclear radiation dose prediction value at each moment and the reconstruction sequence , where the error value is defined as: ； The error value is compared with the set threshold Compare, if there is , then mark this moment as an abnormal point; in is a sequence of reconstruction error values ​​arranged in time order; S64, abnormal label generation unit, used to number the moments marked as abnormal Convert to abnormal event label set and sends the collection to the system control module; in It is the abnormal event label converted from the abnormal moment number.

Images